AI came to 26 NHS stroke units. Thrombectomy doubled — and rose at the other 81 too

Most published evaluations of clinical artificial intelligence measure whether the software can read an image. Very few measure whether anything different happened to the patient afterwards.

A prospective observational study published in The Lancet Digital Health on 2 December did the second thing, at national scale. Researchers used England's national stroke audit registry — the Sentinel Stroke National Audit Programme — to look at every patient admitted to an NHS hospital with a primary diagnosis of stroke between 1 January 2019 and 31 December 2023 [s1]. That is 452,952 patients across all 107 NHS hospitals in England that admit acute stroke [s1].

Over that window, 26 of those hospitals — six comprehensive stroke centres and 20 primary stroke centres — systematically implemented AI stroke imaging software, Brainomix 360 Stroke [s1]. The question was whether the rate of endovascular thrombectomy, the mechanical removal of a clot from a large blood vessel, changed more at those sites than elsewhere.

The numbers

At the 26 evaluation sites, thrombectomy was performed on 2.3% of stroke patients before implementation (376 of 15,969) and 4.6% after (751 of 15,428) — a relative increase of 100% [s1].

At the hospitals that did not implement the software, the rate went from 1.6% (1,431 of 88,712) to 2.6% (2,410 of 89,900), a relative increase of 62.5% [s1].

Both groups improved. The evaluation sites improved more. The odds ratio for the interaction between site and time period was 1.24 (95% CI 1.08–1.43; p=0.0026) [s1].

At the level of individual patients, and restricted to the 71,017 ischaemic stroke patients at evaluation sites for whom patient-level data were available, use of the AI software was associated with an increased likelihood of receiving thrombectomy (odds ratio 1.57, 95% CI 1.33–1.86; p<0.0001) compared with patients whose imaging was reviewed without it [s1].

Why the control group matters more than the headline

The eye goes to "rates doubled." The informative comparison is against the 62.5% relative increase at sites with no AI at all [s1].

Thrombectomy provision in England was expanding across the whole study period for reasons unrelated to software — service commissioning, workforce, referral pathways, and the accumulating trial evidence that made the procedure standard of care for large vessel occlusion [s1]. A before-and-after design at the intervention sites alone would have attributed all of that to the AI. The national comparison is what separates the two, and what it leaves is an interaction odds ratio of 1.24 — real, statistically significant, and considerably smaller than "doubled" [s1].

What the design can and cannot establish

This is a prospective observational study, not a randomised trial [s1]. Hospitals were not allocated to implement AI at random; they chose to, or were selected to, and the study reports that the 26 sites included six comprehensive stroke centres [s1]. Hospitals that adopt new imaging software during a national rollout are plausibly also hospitals investing in stroke pathways more generally — new consultant appointments, faster transfer protocols, better out-of-hours cover. None of those would show up as AI, and all of them would push thrombectomy rates the same way.

The patient-level comparison has a related problem. Cases where the software was used and cases where it was not are not exchangeable; a clinician's decision to run the AI is itself informative about the case. An odds ratio of 1.57 in that comparison is consistent with the software changing decisions, and it is also consistent with the software being used more often on the patients who were already likelier to be treated.

The study was funded by an AI in Health and Care Award from NHS England's Accelerated Access Collaborative [s1]. That is the same programme that paid for the deployment being evaluated, which is common in health-service evaluation and is worth stating plainly.

The context: a field that has trouble recommending anything

Stroke AI is unusual in having an outcome study at all. The more typical position is the one the American Gastroenterological Association reached in March on a different imaging AI — computer-aided polyp detection during colonoscopy — where a guideline panel concluded that no recommendation could be made for or against its use, citing very low certainty of evidence on the outcomes that matter [s2].

That panel's modelled estimates are a useful illustration of why. It projected 11 fewer colorectal cancers and two fewer colorectal cancer deaths per 10,000 people, against 635 more surveillance colonoscopies per 10,000 [s2]. The detection improvement was not in doubt — an 8% increase in adenoma detection rate (95% CI 6–10%) [s2]. What was in doubt was whether the downstream ledger came out ahead.

The English stroke study is measuring a step further along that chain than most AI evaluations reach: not detection accuracy but treatment delivered. It still stops short of the end of the chain, which is disability and death.

What to watch

Whether thrombectomy rates at the evaluation sites hold their gain after the implementation period, and whether any analysis links the treatment increase to functional outcomes at 90 days — the measure stroke medicine actually uses. Rates of a procedure are a proxy for benefit only if the additional patients treated were patients who should have been.

This article is informational and is not medical advice.

Sources

• [s1] Nagaratnam K, Neuhaus AA, Fensome L, et al. "Artificial intelligence imaging decision support for acute stroke treatment in England: a prospective observational study." The Lancet Digital Health, published online 2 December 2025. https://doi.org/10.1016/j.landig.2025.100927
• [s2] Sultan S, Shung DL, Kolb JM, et al. "AGA Living Clinical Practice Guideline on Computer-Aided Detection-Assisted Colonoscopy." Gastroenterology, 168(4), published online 20 March 2025. https://doi.org/10.1053/j.gastro.2025.01.002